深度学习辅助测量强散射涡旋光束拓扑荷数 下载: 624次
1 引言
涡旋光是一种具有螺旋相位波前的奇异光束,其螺旋相位可表示为
然而,在实际应用中还存在一个不容忽视的因素,即湍流或雾霾等对光信号的散射效应。携带信息的涡旋光束经过强散射介质后,由于多重散射的影响,故其相位和振幅受到调制会形成随机散斑。此时,采用传统的干涉或者衍射的方式难以测量其拓扑荷数,致使获取涡旋光所携带的信息成为挑战。因此,仅在非散射环境下测量涡旋光束的拓扑荷数是不够的。Salla等[18]利用偏振和轨道角动量的不可分态,基于对角偏振成分的干涉条纹分瓣数量实现对散射后涡旋光的OAM测量。Singh等[19]提出基于两点强度关联从散斑中恢复波前,从而获知涡旋光的OAM。Gong等[20]提出利用散斑关联散射矩阵,从强散射光场中恢复涡旋光光场的振幅和相位信息,证明了经过散射后的涡旋光中依然存在漩涡状态,该研究促进了涡旋光作为信息载体的应用。然而,上述方法需要多次测量[18-19]或者需要预先测量散射介质的散射矩阵。最近,Chen等[21]基于涡旋光散斑关联的傅里叶变换,实现了从散斑中单次非成像地测量涡旋光拓扑荷数,但这种方法需要额外的参考光路,且对光路的稳定性要求较高。
深度神经网络能够有效地提取数据的内在特征[22],用于图像分类和识别[23]。近年来,卷积神经网络也被用于识别OAM模式和矫正大气湍流引起的相位畸变[24-30]。深度神经网络的引入一定程度上缓解了光通信中湍流引起的干扰。深度学习的方法不需要参考光,相比干涉全息的方法对光路稳定性的要求较低。卷积神经网络具有平移不变性,即使探测器相对光轴发生横向位移也不影响其识别结果[29]。然而,鲜有基于深度学习对涡旋光经过强散射介质的相关研究。因此,本文使用深度学习神经网络,在强散射介质后实现涡旋光拓扑荷数的测量,这种方法对强散射环境中OAM通信的应用具有重要价值。
2 理论分析
图 1. 不同拓扑荷数下的涡旋相位分布和振幅分布。(a)~(c)涡旋相位分布;(d)~(f)振幅分布
Fig. 1. Vortex phase distribution and amplitude distribution under different topological charges. (a)-(c) Vortex phase distribution;
当光经过弱散射介质时,散射介质中主要传播的光为弹道光和蛇形光子,这两者含有入射光所携带的信息,此时只能看见物体的大致轮廓,无法看清细节。强散射介质如生物组织、毛玻璃等对光的多重散射作用更强,在强散射介质中主要是漫散射光子,此时在出射端无法直接获得入射光所携带的原有信息,成像时完全无法看清物体的形貌。
图 2. 涡旋光经过强、弱散射介质后形成的散斑。(a)(d)涡旋光的振幅分布图;(b)(e)涡旋光经过强散射介质后形成的散斑图; 涡旋光经过弱散射介质后形成的散斑图
Fig. 2. Speckles formed by vortex beam passing through strong and weak scattering media. (a)(d) Amplitude distribution patterns of vortex beam; (b)(e) speckle patterns after vortex beam passing through strong scattering medium; (c)(f) speckle patterns after vortex beam passing through weak scattering medium
3 实验光路和数据采集
如
图 3. 涡旋光经过强散射介质的实验光路图
Fig. 3. Experimental optical path diagram of vortex beam passing through strong scattering medium
实验采集的数据集包含了拓扑荷数为±1,
图 4. 不同拓扑荷数的涡旋光经过散射介质后形成的散斑图
Fig. 4. Speckle patterns formed by vortex beams with different topological charges passing through scattering medium
散斑是由激光经过散射介质颗粒散射后造成光线的光程在波长尺寸上各不相同造成的。激光经过散射介质形成的散斑场中存在光强为零的暗点,暗点处的相位不能确定,因此这些暗点被称为相位涡旋。散斑场的相位涡旋与涡旋光的相位息息相关,实验研究表明,涡旋光的散斑场中相位涡旋分布与拓扑荷数有关,拓扑荷数越大,对应散斑场中的相位涡旋的密度就越大[31]。Acevedo等[32]基于菲涅耳衍射模型,描述了涡旋光和完美涡旋光散射产生的随机场的空间相干函数,利用空间相干函数,证明了散斑颗粒大小对涡旋光拓扑荷数的依赖性。以上结论说明不同拓扑荷数的涡旋光形成的散斑具有独特性。
由于相邻拓扑荷数涡旋光散射形成的散斑颗粒尺寸区别较小,为便于观察,选取拓扑荷数分别为1和10的涡旋光形成的散斑图进行对比。在两张散斑图的相同位置截取相同区域,对比该区域中光强最大的散斑颗粒的尺寸。
图 5. 不同拓扑荷数的涡旋光的散斑图和局部放大图。(a)l=1;(b) l=10
Fig. 5. Speckle pattern and local magnification of vortex beam with different topological charges. (a) l=1; (b) l=10
4 神经网络结构和实验结果分析
本文采用卷积神经网络LeNet-5[33]对涡旋光经过散射介质后形成的散斑场进行模式分类。LeNet-5网络最初作为手写体字符号识别的高效神经网络,该网络结构一共有7层,分别为卷积层、池化层、卷积层、池化层和3个线性层,网络结构如
由于本文输入到神经网络中的图片是散斑图,图片结构相对手写体字符更加复杂,为让网络能够实现非线性分类,神经网络的每个卷积层后均加入了ReLU激活函数。输入散斑灰度图的大小为256 pixel
表 1. 神经网络各层参数设置
Table 1. Parameter setting of each layer of neural network
|
通过实验采集得到了20类不同拓扑荷数的涡旋光对应的散斑图,其中每一类有100张,共计得到2000张散斑图。将每一类散斑图的80%(数量占比)作为训练集,10%作为验证集,剩下的10%作为测试集。在中央处理器(CPU,i7-10750H)环境下对散斑图进行训练,输入到神经网络中的图片的分辨率为256 pixel×256 pixel,训练过程中损失值的变化情况如
图 7. 10个周期内训练损失与验证损失的曲线图
Fig. 7. Curves of training loss and validation loss in 10 epochs
为进一步获取训练过程中神经网络对训练集和验证集的学习情况,引入了准确度的变化情况。如
图 8. 10个周期内训练准确度与验证准确度的曲线图
Fig. 8. Curves of training accuracy and validation accuracy in 10 epochs
利用深度学习训练大量数据的目的是得到一个具有泛化能力强且质量好的网络模型,再利用得到的模型对未参与训练的散斑图进行计算,且最终能够精确地测量散斑对应的拓扑荷数。
5 结论
涡旋光束经过强散射介质后会形成杂乱无序的散斑。与传统方式相比,利用深度学习的方法能够更高效且准确地测量经过非均匀介质涡旋光的拓扑荷数。与弱散射介质相比,强散射介质对光束所携带的信息的破坏更加严重。当利用OAM光束实现光通信时,强散射介质会破坏OAM通道之间的正交性,进而导致较高的误码率。基于深度学习,实现了强散射后涡旋光拓扑荷数的测量,且准确率达到100%。研究结果对于强散射环境下的涡旋光通信具有重要意义。
[1] Allen L, Beijersbergen M W, Spreeuw R J C, et al. Orbital angular momentum of light and the transformation of Laguerre-Gaussian laser modes[J]. Physical Review A, 1992, 45(11): 8185-8189.
[2] Padgett M, Bowman R. Tweezers with a twist[J]. Nature Photonics, 2011, 5(6): 343-348.
[3] Wang J. Twisted optical communications using orbital angular momentum[J]. Science China Physics, Mechanics & Astronomy, 2018, 62(3): 034201.
[4] Wang X L, Luo Y H, Huang H L, et al. 18-qubit entanglement with six photons' three degrees of freedom[J]. Physical Review Letters, 2018, 120(26): 260502.
[5] Willner A E, Huang H, Yan Y, et al. Optical communications using orbital angular momentum beams[J]. Advances in Optics and Photonics, 2015, 7(1): 66-106.
[6] Willner A E, Ren Y X, Xie G D, et al. Recent advances in high-capacity free-space optical and radio-frequency communications using orbital angular momentum multiplexing[J]. Philosophical Transactions of the Royal Society A, 2017, 375(2087): 20150439.
[7] Chen L, Lei J, Romero J. Quantum digital spiral imaging[J]. Light: Science & Applications, 2014, 3(3): e153.
[8] Gan W W, Chen L F, Liu Y Q. Research on optical encryption system based on unequal modulus decomposition and polarized vortex optical holography[J]. Optics Communications, 2021, 482: 126609.
[9] Sui L S, Zhou B, Ning X J, et al. Optical multiple-image encryption based on the chaotic structured phase masks under the illumination of a vortex beam in the gyrator domain[J]. Optics Express, 2016, 24(1): 499-515.
[11] Zhang W H, Qi Q Q, Zhou J, et al. Mimicking Faraday rotation to sort the orbital angular momentum of light[J]. Physical Review Letters, 2014, 112(15): 153601.
[12] Ghai D P, Vyas S, Senthilkumaran P, et al. Detection of phase singularity using a lateral shear interferometer[J]. Optics and Lasers in Engineering, 2008, 46(6): 419-423.
[13] Sztul H I, Alfano R R. Double-slit interference with Laguerre-Gaussian beams[J]. Optics Letters, 2006, 31(7): 999-1001.
[14] Guo C S, Lu L L, Wang H T. Characterizing topological charge of optical vortices by using an annular aperture[J]. Optics Letters, 2009, 34(23): 3686-3688.
[15] Vaity P, Banerji J, Singh R P. Measuring the topological charge of an optical vortex by using a tilted convex lens[J]. Physics Letters A, 2013, 377(15): 1154-1156.
[16] Narag J P C, Hermosa N. Probing higher orbital angular momentum of Laguerre-Gaussian beams via diffraction through a translated single slit[J]. Physical Review Applied, 2019, 11(5): 054025.
[17] Lu J N, Cao C Y, Zhu Z Q, et al. Flexible measurement of high-order optical orbital angular momentum with a variable cylindrical lens pair[J]. Applied Physics Letters, 2020, 116(20): 201105.
[18] Salla G R, Perumangattu C, Prabhakar S, et al. Recovering the vorticity of a light beam after scattering[J]. Applied Physics Letters, 2015, 107(2): 021104.
[19] Singh R K, Vinu V R, Anandraj S M. Recovery of complex valued objects from two-point intensity correlation measurement[J]. Applied Physics Letters, 2014, 104(11): 111108.
[20] Gong L, Zhao Q, Zhang H, et al. Optical orbital-angular-momentum-multiplexed data transmission under high scattering[J]. Light: Science & Applications, 2019, 8: 27.
[22] LeCun Y, Bengio Y, Hinton G. Deep learning[J]. Nature, 2015, 521(7553): 436-444.
[23] Krizhevsky A, Sutskever I, Hinton G E. ImageNet classification with deep convolutional neural networks[J]. Communications of the ACM, 2017, 60(6): 84-90.
[24] Liu J M, Wang P P, Zhang X K, et al. Deep learning based atmospheric turbulence compensation for orbital angular momentum beam distortion and communication[J]. Optics Express, 2019, 27(12): 16671-16688.
[25] Mao Z X, Yu H Y, Xia M, et al. Broad bandwidth and highly efficient recognition of optical vortex modes achieved by the neural-network approach[J]. Physical Review Applied, 2020, 13(3): 034063.
[26] Lohani S, Glasser R T. Turbulence correction with artificial neural networks[J]. Optics Letters, 2018, 43(11): 2611-2614.
[27] Lohani S, Knutson E M, O'Donnell M, et al. On the use of deep neural networks in optical communications[J]. Applied Optics, 2018, 57(15): 4180-4190.
[28] Doster T, Watnik A T. Machine learning approach to OAM beam demultiplexing via convolutional neural networks[J]. Applied Optics, 2017, 56(12): 3386-3396.
[29] Liu Z W, Yan S, Liu H G, et al. Superhigh-resolution recognition of optical vortex modes assisted by a deep-learning method[J]. Physical Review Letters, 2019, 123(18): 183902.
[30] Na Y, Ko D K. Deep-learning-based high-resolution recognition of fractional-spatial-mode-encoded data for free-space optical communications[J]. Scientific Reports, 2021, 11: 2678.
[31] 刘曼. 涡旋光束形成的散斑场光强和相位的分布特性[J]. 光学学报, 2014, 34(11): 1126001.
[32] Acevedo C H, Torres-Moreno Y, Dogariu A. Spatial intensity correlations of a vortex beam and a perfect optical vortex beam[J]. Journal of the Optical Society of America A, 2019, 36(4): 518-525.
Article Outline
刘雪莲, 陈旭东, 林志立, 刘卉, 朱香渝, 张晓雪. 深度学习辅助测量强散射涡旋光束拓扑荷数[J]. 光学学报, 2022, 42(14): 1426001. Xuelian Liu, Xudong Chen, Zhili Lin, Hui Liu, Xiangyu Zhu, Xiaoxue Zhang. Deep-Learning-Assisted Detection For Topological Charges of Vortex Beams Through Strong Scattering Medium[J]. Acta Optica Sinica, 2022, 42(14): 1426001.